Deep Conditional Gaussian Mixture Model for Constrained Clustering
Authors: Laura Manduchi, Kieran Chin-Cheong, Holger Michel, Sven Wellmann, Julia Vogt
NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We provide extensive experiments to demonstrate that DC-GMM shows superior clustering performances and robustness compared to state-of-the-art deep constrained clustering methods on a wide range of data sets. We further demonstrate the usefulness of our approach on two challenging real-world applications. |
| Researcher Affiliation | Academia | Laura Manduchi Department of Computer Science ETH Zürich laura.manduchi@inf.ethz.ch Kieran Chin-Cheong Department of Computer Science ETH Zürich kieran.chincheong@inf.ethz.ch Holger Michel Department of Neonatology University Children s Hospital Regensburg (KUNO) University of Regensburg, Germany holger.michel@barmherzige-regensburg.de Sven Wellmann Department of Neonatology University Children s Hospital Regensburg (KUNO) University of Regensburg, Germany sven.wellmann@klinik.uni-regensburg.de Julia E. Vogt Department of Computer Science ETH Zürich julia.vogt@inf.ethz.ch |
| Pseudocode | No | The paper describes steps and equations for the model but does not include structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | The code is available in a Git Hub repository: https://github.com/lauramanduchi/DC-GMM. |
| Open Datasets | Yes | MNIST (Le Cun et al., 2010), Fashion MNIST (Xiao et al., 2017), Reuters (Xie et al., 2016) and STL-10 (Coates et al., 2011) (see Appendix A). ... The Heart Echo data is not available due to medical confidentiality. |
| Dataset Splits | No | The paper mentions "Each data set is divided into training and test sets" but does not specify validation splits or exact percentages/counts for each split needed for reproduction. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., exact GPU/CPU models, memory amounts) used for running its experiments. |
| Software Dependencies | No | The paper does not provide specific ancillary software details (e.g., library or solver names with version numbers like Python 3.8, PyTorch 1.9) needed to replicate the experiment. |
| Experiment Setup | Yes | The VAE is pretrained for 10 epochs while the DEC-based baselines need a more complex layer-wise pretraining of the autoencoder which involves 50 epochs of pretraining for each layer and 100 epochs of pretraining as finetuning. Each data set is divided into training and test sets, and all the reported results are computed on the latter. We employed the same hyper-parameters for all data sets, see Appendix F.1 for details. The pairwise constraints are chosen randomly within the training set by sampling two data points and assigning a must-link if they have the same label and a cannot-link otherwise. Unless stated otherwise, the values of |Wi,j| are set to 104 for all data sets, and 6000 pairwise constraints are used for both our model and the constrained clustering baselines. |